aitech-eks-pub/cw/09_sequence_labeling.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![Logo 1](https://git.wmi.amu.edu.pl/AITech/Szablon/raw/branch/master/Logotyp_AITech1.jpg)\n",
    "<div class=\"alert alert-block alert-info\">\n",
    "<h1> Ekstrakcja informacji </h1>\n",
    "<h2> 9. <i>Sequence labeling</i>  [ćwiczenia]</h2> \n",
    "<h3> Jakub Pokrywka (2021)</h3>\n",
    "</div>\n",
    "\n",
    "![Logo 2](https://git.wmi.amu.edu.pl/AITech/Szablon/raw/branch/master/Logotyp_AITech2.jpg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Klasyfikacja wieloklasowa i sequence labelling"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import gensim\n",
    "import torch\n",
    "import pandas as pd\n",
    "import seaborn as sns\n",
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "from datasets import load_dataset\n",
    "from torchtext.vocab import Vocab\n",
    "from collections import Counter\n",
    "\n",
    "from sklearn.datasets import fetch_20newsgroups\n",
    "# https://scikit-learn.org/0.19/datasets/twenty_newsgroups.html\n",
    "\n",
    "from sklearn.feature_extraction.text import TfidfVectorizer\n",
    "from sklearn.metrics import accuracy_score"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Klasyfikacja"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Klasfikacja binarna-  2 klasy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "CATEGORIES = ['soc.religion.christian', 'alt.atheism']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "newsgroups = fetch_20newsgroups(categories=CATEGORIES)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "X =  newsgroups['data']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y = newsgroups['target']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y_names = newsgroups['target_names']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "X[0:1]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "Y_names"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "del CATEGORIES, newsgroups, X, Y, Y_names"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### klasyfikacja wieloklasowa"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "newsgroups_train_dev = fetch_20newsgroups(subset = 'train')\n",
    "newsgroups_test = fetch_20newsgroups(subset = 'test')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "newsgroups_train_dev_text = newsgroups_train_dev['data']\n",
    "newsgroups_test_text = newsgroups_test['data']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y_train_dev = newsgroups_train_dev['target']\n",
    "Y_test = newsgroups_test['target']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "newsgroups_train_text, newsgroups_dev_text, Y_train, Y_dev = train_test_split(newsgroups_train_dev_text, Y_train_dev, random_state=42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y_names = newsgroups_train_dev['target_names']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y_train_dev"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y_names"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**Jaki baseline?**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "pd.value_counts(Y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "accuracy_score(Y_test, np.ones_like(Y_test) * 10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "\n",
    "**Pytanie** - w jaki sposób stworzyć taki klasyfikator na podstawie tylko wiedzy z poprzednich ćwiczeń?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Zadanie - stworzyć klasyfikator regresji logistycznej one vs rest na podstawie tfdif. TFIDF powinien mieć słownik o wielkości 10000\n",
    "\n",
    "https://scikit-learn.org/stable/modules/generated/sklearn.multiclass.OneVsRestClassifier.html\n",
    "https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html\n",
    "https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.TfidfVectorizer.html"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.multiclass import OneVsRestClassifier\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "from sklearn.feature_extraction.text import TfidfVectorizer"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Podejście softmax na tfidif"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**Zadanie** Na podstawie poprzednich zajęć stworzyć sieć w pytorch bez warstw ukrytych, z jedną warstwą *output* z funkcją softmax (bez trenowania i ewaluacji sieci)\n",
    "\n",
    "Użyć https://pytorch.org/docs/stable/generated/torch.nn.Softmax.html?highlight=softmax"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "class NeuralNetworkModel(torch.nn.Module):\n",
    "\n",
    "    pass"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "OUTPUT_SIZE = len(Y_names)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "nn_model = NeuralNetworkModel(FEAUTERES, OUTPUT_SIZE)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "nn_model(torch.Tensor(X_train[0:3].astype(np.float32).todense()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "BATCH_SIZE = 5"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "criterion = torch.nn.NLLLoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = torch.optim.SGD(nn_model.parameters(), lr = 0.2)\n",
    "#optimizer = torch.optim.Adam(nn_model.parameters())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_loss_acc(model, X_dataset, Y_dataset):\n",
    "    loss_score = 0\n",
    "    acc_score = 0\n",
    "    items_total = 0\n",
    "    model.eval()\n",
    "    for i in range(0, Y_dataset.shape[0], BATCH_SIZE):\n",
    "        X = X_dataset[i:i+BATCH_SIZE]\n",
    "        X = torch.tensor(X.astype(np.float32).todense())\n",
    "        Y = Y_dataset[i:i+BATCH_SIZE]\n",
    "        Y = torch.tensor(Y)\n",
    "        Y_predictions = model(X)\n",
    "        acc_score += torch.sum(torch.argmax(Y_predictions,dim=1) == Y).item()\n",
    "        items_total += Y.shape[0] \n",
    "\n",
    "        loss = criterion(Y_predictions, Y)\n",
    "\n",
    "        loss_score += loss.item() * Y.shape[0] \n",
    "    return (loss_score / items_total), (acc_score / items_total)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "for epoch in range(5):\n",
    "    loss_score = 0\n",
    "    acc_score = 0\n",
    "    items_total = 0\n",
    "    nn_model.train()\n",
    "    for i in range(0, Y_train.shape[0], BATCH_SIZE):\n",
    "        X = X_train[i:i+BATCH_SIZE]\n",
    "        X = torch.tensor(X.astype(np.float32).todense())\n",
    "        Y = Y_train[i:i+BATCH_SIZE]\n",
    "\n",
    "        Y = torch.tensor(Y)\n",
    "        Y_predictions = nn_model(X)\n",
    "        acc_score += torch.sum(torch.argmax(Y_predictions,dim=1) == Y).item()\n",
    "        items_total += Y.shape[0] \n",
    "\n",
    "        optimizer.zero_grad()\n",
    "        loss = criterion(Y_predictions, Y)\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "\n",
    "\n",
    "        loss_score += loss.item() * Y.shape[0]\n",
    "\n",
    "    \n",
    "    display(epoch)\n",
    "    display(get_loss_acc(nn_model, X_train, Y_train))\n",
    "    display(get_loss_acc(nn_model, X_dev, Y_dev))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "X.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "newsgroups_train_text"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Podejście softmax z embeddingami na przykładzie NER"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# !pip install torchtext\n",
    "# !pip install datasets"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "https://www.aclweb.org/anthology/W03-0419.pdf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "dataset = load_dataset(\"conll2003\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def build_vocab(dataset):\n",
    "    counter = Counter()\n",
    "    for document in dataset:\n",
    "        counter.update(document)\n",
    "    return Vocab(counter, specials=['<unk>', '<pad>', '<bos>', '<eos>'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "vocab = build_vocab(dataset['train']['tokens'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset['train']['tokens']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "len(vocab.itos)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "vocab['on']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def data_process(dt):\n",
    "    return [ torch.tensor([vocab['<bos>']] +[vocab[token]  for token in  document ] + [vocab['<eos>']], dtype = torch.long) for document in dt]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def labels_process(dt):\n",
    "    return [ torch.tensor([0] + document + [0], dtype = torch.long) for document in dt]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_tokens_ids = data_process(dataset['train']['tokens'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_tokens_ids = data_process(dataset['test']['tokens'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "train_labels = labels_process(dataset['train']['ner_tags'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_labels = labels_process(dataset['test']['ner_tags'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_tokens_ids[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "max([max(x) for x in dataset['train']['ner_tags'] ])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "class NERModel(torch.nn.Module):\n",
    "\n",
    "    def __init__(self,):\n",
    "        super(NERModel, self).__init__()\n",
    "        self.emb = torch.nn.Embedding(23627,200)\n",
    "        self.fc1 = torch.nn.Linear(600,9)\n",
    "        #self.softmax = torch.nn.Softmax(dim=0)\n",
    "        # nie trzeba, bo używamy https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html\n",
    "        # jako kryterium\n",
    "        \n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.emb(x)\n",
    "        x = x.reshape(600) \n",
    "        x = self.fc1(x)\n",
    "        #x = self.softmax(x)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_tokens_ids[0][1:4]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ner_model = NERModel()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ner_model(train_tokens_ids[0][1:4])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "criterion = torch.nn.CrossEntropyLoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = torch.optim.Adam(ner_model.parameters())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "len(train_labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "for epoch in range(2):\n",
    "    loss_score = 0\n",
    "    acc_score = 0\n",
    "    prec_score = 0\n",
    "    selected_items = 0\n",
    "    recall_score = 0\n",
    "    relevant_items = 0\n",
    "    items_total = 0\n",
    "    nn_model.train()\n",
    "    #for i in range(len(train_labels)):\n",
    "    for i in range(100):\n",
    "        for j in range(1, len(train_labels[i]) - 1):\n",
    "    \n",
    "            X = train_tokens_ids[i][j-1: j+2]\n",
    "            Y = train_labels[i][j: j+1]\n",
    "\n",
    "            Y_predictions = ner_model(X)\n",
    "            \n",
    "            \n",
    "            acc_score += int(torch.argmax(Y_predictions) == Y)\n",
    "            \n",
    "            if torch.argmax(Y_predictions) != 0:\n",
    "                selected_items +=1\n",
    "            if  torch.argmax(Y_predictions) != 0 and torch.argmax(Y_predictions) == Y.item():\n",
    "                prec_score += 1\n",
    "            \n",
    "            if  Y.item() != 0:\n",
    "                relevant_items +=1\n",
    "            if  Y.item() != 0 and torch.argmax(Y_predictions) == Y.item():\n",
    "                recall_score += 1\n",
    "            \n",
    "            items_total += 1\n",
    "\n",
    "            \n",
    "            optimizer.zero_grad()\n",
    "            loss = criterion(Y_predictions.unsqueeze(0), Y)\n",
    "            loss.backward()\n",
    "            optimizer.step()\n",
    "\n",
    "\n",
    "            loss_score += loss.item() \n",
    "    \n",
    "    precision = prec_score / selected_items\n",
    "    recall = recall_score / relevant_items\n",
    "    f1_score = (2*precision * recall) / (precision + recall)\n",
    "    display('epoch: ', epoch)\n",
    "    display('loss: ', loss_score / items_total)\n",
    "    display('acc: ', acc_score / items_total)\n",
    "    display('prec: ', precision)\n",
    "    display('recall: : ', recall)\n",
    "    display('f1: ', f1_score)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "loss_score = 0\n",
    "acc_score = 0\n",
    "prec_score = 0\n",
    "selected_items = 0\n",
    "recall_score = 0\n",
    "relevant_items = 0\n",
    "items_total = 0\n",
    "nn_model.eval()\n",
    "for i in range(100):\n",
    "#for i in range(len(test_labels)):\n",
    "    for j in range(1, len(test_labels[i]) - 1):\n",
    "\n",
    "        X = test_tokens_ids[i][j-1: j+2]\n",
    "        Y = test_labels[i][j: j+1]\n",
    "\n",
    "        Y_predictions = ner_model(X)\n",
    "\n",
    "\n",
    "        acc_score += int(torch.argmax(Y_predictions) == Y)\n",
    "\n",
    "        if torch.argmax(Y_predictions) != 0:\n",
    "            selected_items +=1\n",
    "        if  torch.argmax(Y_predictions) != 0 and torch.argmax(Y_predictions) == Y.item():\n",
    "            prec_score += 1\n",
    "\n",
    "        if  Y.item() != 0:\n",
    "            relevant_items +=1\n",
    "        if  Y.item() != 0 and torch.argmax(Y_predictions) == Y.item():\n",
    "            recall_score += 1\n",
    "\n",
    "        items_total += 1\n",
    "\n",
    "\n",
    "        loss = criterion(Y_predictions.unsqueeze(0), Y)\n",
    "\n",
    "\n",
    "\n",
    "        loss_score += loss.item() \n",
    "\n",
    "precision = prec_score / selected_items\n",
    "recall = recall_score / relevant_items\n",
    "f1_score = (2*precision * recall) / (precision + recall)\n",
    "display('loss: ', loss_score / items_total)\n",
    "display('acc: ', acc_score / items_total)\n",
    "display('prec: ', precision)\n",
    "display('recall: : ', recall)\n",
    "display('f1: ', f1_score)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Zadanie domowe\n",
    "\n",
    "- sklonować repozytorium https://git.wmi.amu.edu.pl/kubapok/en-ner-conll-2003\n",
    "- stworzyć klasyfikator bazujący na sieci neuronowej feed forward w pytorchu (można bazować na tym jupyterze lub nie).\n",
    "- klasyfikator powinien obejmować dodatkowe cechy (np. długość wyrazu, czy wyraz zaczyna się od wielkiej litery, stemmming słowa, czy zawiera cyfrę)\n",
    "- stworzyć predykcje w plikach dev-0/out.tsv oraz test-A/out.tsv\n",
    "- wynik fscore sprawdzony za pomocą narzędzia geval (patrz poprzednie zadanie) powinien wynosić conajmniej 0.60\n",
    "- proszę umieścić predykcję oraz skrypty generujące (w postaci tekstowej a nie jupyter) w repo, a w MS TEAMS umieścić link do swojego repo\n",
    "termin 08.06, 80 punktów\n"
   ]
  }
 ],
 "metadata": {
  "author": "Jakub Pokrywka",
  "email": "kubapok@wmi.amu.edu.pl",
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "lang": "pl",
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.3"
  },
  "subtitle": "9.Sequence labeling[ćwiczenia]",
  "title": "Ekstrakcja informacji",
  "year": "2021"
 },
 "nbformat": 4,
 "nbformat_minor": 4
}